This is the race scenario based on call trace we captured which cause
the dentry leak.
CPU 0 CPU 1
ovl_set_redirect lookup_fast
ovl_get_redirect __d_lookup
dget_dlock
//no lock protection here spin_lock(&dentry->d_lock)
dentry->d_lockref.count++ dentry->d_lockref.count++
If we use dget_parent instead, we may have this race.
CPU 0 CPU 1
ovl_set_redirect lookup_fast
ovl_get_redirect __d_lookup
dget_parent
raw_seqcount_begin(&dentry->d_seq) spin_lock(&dentry->d_lock)
lockref_get_not_zero(&ret->d_lockref) dentry->d_lockref.count++
On 20/12/21 下午8:11, Al Viro wrote:
On Mon, Dec 21, 2020 at 07:14:44PM +0800, Joseph Qi wrote:
Hi Viro,
On 12/21/20 2:26 PM, Al Viro wrote:
On Sun, Dec 20, 2020 at 08:09:27PM +0800, Liangyan wrote:
+++ b/fs/overlayfs/dir.c
@@ -973,6 +973,7 @@ static char *ovl_get_redirect(struct dentry *dentry, bool abs_redirect)
for (d = dget(dentry); !IS_ROOT(d);) {
const char *name;
int thislen;
+ struct dentry *parent = NULL;
spin_lock(&d->d_lock);
name = ovl_dentry_get_redirect(d);
@@ -992,7 +993,22 @@ static char *ovl_get_redirect(struct dentry *dentry, bool abs_redirect)
buflen -= thislen;
memcpy(&buf[buflen], name, thislen);
- tmp = dget_dlock(d->d_parent);
+ parent = d->d_parent;
+ if (unlikely(!spin_trylock(&parent->d_lock))) {
+ rcu_read_lock();
+ spin_unlock(&d->d_lock);
+again:
+ parent = READ_ONCE(d->d_parent);
+ spin_lock(&parent->d_lock);
+ if (unlikely(parent != d->d_parent)) {
+ spin_unlock(&parent->d_lock);
+ goto again;
+ }
+ rcu_read_unlock();
+ spin_lock_nested(&d->d_lock, DENTRY_D_LOCK_NESTED);
+ }
+ tmp = dget_dlock(parent);
+ spin_unlock(&parent->d_lock);
spin_unlock(&d->d_lock);
Yecchhhh.... What's wrong with just doing
spin_unlock(&d->d_lock);
parent = dget_parent(d);
dput(d);
d = parent;
instead of that?
Now race happens on non-RCU path in lookup_fast(), I'm afraid d_seq can
not close the race window.
Explain, please. What exactly are you observing?
